Biologically inspired sensor fusion for real-time wind gust estimation in autonomous UAV navigation Reliability and accuracy of navigation in flying drones are one of the key challenges that must be solved in autonomous applications. Accurate knowledge of position, attitude, and velocity is a critical input for drones operating in cluttered and challenging environments such as cities, forests or mountain terrain. The possibility of (Global Navigation Satellite System) GNSS failure due to signal obstruction or external interference can cause complete failure of a drone navigation system (Figure 1) which would be unacceptable for any future beyond visual line of site commercial applications. > Figure 1: Drone position error growing fast post GNSS failure [1] This challenge can be solved if an accurate enough alternative to the GNSS navigation system is available as complementary in such conditions. Current state of the art solution at TOPO laboratory is the
application of Vehicle Dynamics Model based navigation [1]. Its application has shown significant improvement in critical situations related to GNSS outage (Figure 2). Figure 2: Improvement in navigation error post GNSS outage via VDM based navigation As seen in (Figure 2) about 200 seconds post GNSS failure the accumulated error in knowing the drone location is more than 2 km! With the use of VDM based navigation as complimentary, the error is reduced to around 50 m (with the assumption of no gusts or regular wind is present)! A key challenge for the implementation of this innovation are effects related to wind/gust disturbances. Unexpected wind gust variation can cause large navigation error (Figure 3) that could make the VDM system drift 10 s or even 100 eds of meters. On the contrary, an accurate estimation of the wind effect could even further improve the VDM based navigation to accuracies similar to the GNSS one. Those effects are difficult to observe and estimate by classical means (Inertial sensor inference or pitot tube 1 Dimensional wind velocity measurements).
Figure 3: Cause and effect of wind gust to fixed wing UAV [2], [3] A new research direction has initiated within the TOPO laboratory drawing inspiration from nature for the solution of this problem. Wind and gust disturbance is managed rather elegantly by birds and nature by fusing information from several sensorial systems. Among this direct aero surface wind direction and strength measurement done by their feathers [2], [3] (Figure 3). Figure 4: Birds airflow and gust sensation via feathers [2], [3], [5]
A several type (heat-flux based, and pressure based) Micro-Electro-Mechanical Systems (MEMS) sensors are being investigated as possible candidates for direct wind effect sensation. Those are envisaged to be integrated directly on the wing/aerodynamic surface. The upcoming research challenges lie in the capability of accurately sensing local wind velocity and/or pressure vectors on discrete measurements points on the wing and relating those to the complete 3D aerodynamic forces profile (Figure 3). Deep Neural Net, Physical model or hybrid Figure 3: Development of model linking discrete measurements with complete aerodynamic (Lift, Drag) forces profile in real time. Numerous practical/experimental as well as modeling and computation challenges are described below. Those may define several upcoming Semester, Master or even Ph.D. related topics. RESEARCH TOPICS/OPEN QUESTIONS Real-time estimation of the total wind gust force on drones center of gravity by several discrete sensor measurements? 1) Experimental (Spring 2019 - wind tunnel currently under refurbishment): a. What is the wind gust profile (thus total force vector) measurement uncertainty achieved at EPFL wind tunnel? Answering this question would provide an absolute reference for any further experiments and modeling research. b. Evaluation of current commercial drone pitot tube wind velocity measurement sensors performance and capabilities for their bias and stochastic errors compensations. This portion of work would set the reference of what can be achieved with current off-theshelf available equipment. c. What is the pointwise measurement uncertainty, sensitivity and practical performance of: state-of-the-art air pressure and beyond state-of-the-art velocity vector measurement sensors (heat-flux based). This experimental work would validate the performance of the novel sensors envisaged to be integrated within the drone skin surface. d. Validation/intercomparing between wind tunnel measurements and future models (based on Computational Fluid Dynamics CFD/deep learning) used to predict aerodynamic force profiles.
2) Modeling (Currently available): a. What is the uncertainty of steady-state CFD simulation (relation to wind tunnel experiments)? b. What is the transient state (gust) uncertainty of CFD simulation (relation to wind tunnel experiments)? c. What can be the uncertainty of a metamodel (Deep Neural Net or other) that can link discrete sensor readings to complete flow profile? Uncertainty in UAV s autonomous navigation due to wind gust velocity vector estimation (semester project Schenker Theophile) Figure 4: Deep Learning based prediction of aerodynamics based on CFD data used for model training [4] A novel approach of accelerating computationally expensive CFD (Computational Fluid Dynamics) simulations is emerging by the application of a Deep Learning Neural Networks based metamodels [4]. For this approach to be applied first input/output database with known uncertainty of the data must be established. In the case of CFD analysis, this would be a database relating discrete sensorial measurements or CFD model readings as input (surface velocity or pressure vectors) and output (a predicted velocity and pressure aerodynamic profiles). Such a database has to be created via CFD modeling with knowledge and understanding of the uncertainty related to the model parameters and CFD model itself. The student that takes up the challenge would have to develop a working parametrized CFD simulation (both steady state and transient) of a 2D standard NACA (National Advisory Committee for Aeronautics https://en.wikipedia.org/wiki/naca_airfoil ) wing model. This simulation would be the backbone for automatized simulation workflow that would generate the required CFD simulation database required for the training of Deep Neural Network that will represent the real aerodynamic behavior. Beyond the student s project (future/continuation work) would be the usage of the available database for the creation of the DNN model. The student that takes up this intriguing challenge should complete under supervision the following (set of mandatory 1-3 and set of complimentary 5) tasks:
1) Compile literature and review it summarised from a given set of papers and tutorials provided as introductory to the challenge. a. General understanding of the challenge. b. Preparation and understanding of the CFD model. c. Uncertainty analysis of a model. 2) Preparation of a parametrized CFD model based on ANSYS Fluent (https://www.ansys.com/products/fluids) and the novel ANSYS discovery live (https://www.ansys.com/products/3d-design/ansys-discovery-live ), software solutions. 3) Study of the variance of the CFD model prediction as a function of the input parameters variance. This would help in a later stage for the estimation of the CFD model s uncertainty once the wind tunnel complex is being made available. 4) Summary of the results in a report/presentation complimentary 5) Definition of the Deep Learning Network (DNN) Input/output database that will serve as a starting point for the next stage - the training and uncertainty estimation of such network classification. a. Literature review on Deep Learning Models architecture related to input/output dimensionality related challenges (time-based series). b. Suggestion for the next stage of the work (the DNN development/training). Key dates:( to be discussed and fixed on a schedule between the student and leading Postdoc, it is expected by the student to design an appropriate working schedule Gantt chart to be evaluated by the Postdoc.) Official start 18.02.2019 Introductorily presentation showing student understanding of the topic (2 weeks following the project start) Pre-finish presentation (3 weeks prior to final agreed date) Official finish: 06.2019
References: [1] M. Khaghani and J. Skaloud, Assessment of VDM-based autonomous navigation of a UAV under operational conditions, Rob. Auton. Syst., vol. 106, no. 106, pp. 152 164, Aug. 2018. [2] A. Mohamed, K. Massey, S. Watkins, and R. Clothier, The attitude control of fixed-wing MAVS in turbulent environments, Prog. Aerosp. Sci., vol. 66, pp. 37 48, 2014. [3] A. Mohamed, S. Watkins, R. Clothier, M. Abdulrahim, K. Massey, and R. Sabatini, Fixedwing MAV attitude stability in atmospheric turbulence - Part 2: Investigating biologicallyinspired sensors, Prog. Aerosp. Sci., vol. 71, pp. 1 13, 2014. [4] N. Umetani and B. Bickel, Learning three-dimensional flow for interactive aerodynamic design, ACM Trans. Graph., vol. 37, no. 4, pp. 1 10, 2018. [5] G. Taylor, M. Bacic, A. Carruthers, J. Gillies, Y. Ozawa, and A. Thomas, Flight Control Mechanisms in Birds of Prey, 45th AIAA Aerosp. Sci. Meet. Exhib., no. January, 2007.